Suitability of Deep Weakly Supervised Learning to detect Acute Ischemic Stroke and Hemorrhagic Infarction Lesions Using Diffusion-weighted Imaging

ObjectivesThe automatic detection of acute ischemic stroke (AIS) and hemorrhagic infarction (HI) based on deep learning could avoid missed diagnosis. The fully supervised learning requires the amount of time and the expertise to manually outline lesions, which limits its applicability. The weakly supervised learning has the potential to reduce the labeling workload. The purpose of this study was to evaluate a weakly supervised method in detection of AIS and HI location using DWI.MethodsWe proposed to adopt weakly supervised learning to spatially-locate AIS lesions by residual neural network (ResNet) and visual geometry group (VGG) network. On an AIS dataset, the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity and precision were calculated. Next, ResNet, which presented superior performance on the AIS dataset, was further applied to an HI dataset.ResultsIn the AIS dataset, the AUCs of ResNet and VGG on identifying image slices with AIS were 0.97 and 0.94, respectively. On spatially-locating the AIS lesions, ResNet provided higher sensitivity and a lower missed diagnosis rate than VGG, especially for pontine AIS lesions. In the HI dataset, the sensitivity of ResNet was 87.73% for AIS detection, and 86.20% for HI detection, respectively.ConclusionsWeakly supervised learning can effectively detect the location of AIS and HI lesions in DWI, which is of paramount importance in avoiding misdiagnosis in clinical scenario.Key pointsThe deep weakly supervised learning can reduce the labeling workload; ResNet can obtain more exact results, especially for pontine AIS lesions; Weakly supervised learning can effectively detect AIS and HI lesions in DWI